[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-researchers-squeeze-more-text-into-ai-context-windows-with-images":10,"sections":35},{"siteName":4,"siteTagline":5,"publisherName":4,"contactEmail":6},"The Revision","Tech news, decoded.","editor@therevision.news",{"gaMeasurementId":8,"adsenseClientId":9},"G-ZW2MV82GYR","ca-pub-8533917693782264",{"article":11},{"id":12,"slug":13,"title":14,"dek":15,"body_md":16,"tags_json":17,"published_at":18,"created_at":19,"updated_at":20,"status":21,"review_note":22,"review_notes":23,"image_url":22,"persona_id":22,"persona_name":22,"section":24,"tags":25,"sources":30,"feedback":34,"feedback_at":22,"cost_usd":34,"total_tokens":34},8602,"researchers-squeeze-more-text-into-ai-context-windows-with-images","Researchers Squeeze More Text Into AI Context Windows With Images","A new method renders text as adaptive-resolution images to shrink AI inputs, beating a rival compression approach without sacrificing accuracy.","A new technique called FocusVTC compresses long documents into images so AI models can read them using far fewer tokens, without the usual loss in accuracy.\n\nResearchers behind the arXiv paper built FocusVTC to fix a known problem with visual text compression, a method that renders text as pictures to shrink input length. Low-resolution renders save tokens but blur small text; high-resolution renders stay legible but waste tokens on parts of the page nobody needs. FocusVTC handles this by starting with a compressed, low-DPI view of a document and then selectively re-rendering only the regions the model's reasoning flags as relevant. To teach it which regions matter, the team built a 29,400-example training set that links reasoning steps to specific page locations, then used supervised fine-tuning plus reinforcement learning to teach the model when to zoom in.\n\nAt matched compression ratios, FocusVTC scored 87.4 versus 57.5 for Glyph, a prior visual-compression approach, on the RULER v1 long-context benchmark, a wide enough gap to suggest adaptive resolution was the missing piece, not just compression. It also ran 2.79 times faster end to end than a plain text pipeline on the MRCR benchmark, while slightly improving scores on general multimodal tests like MMMU and MME rather than trading them away.\n\nIt's still a lab result with self-reported benchmarks and no shipped product, but the idea that pictures can be a cheaper way for machines to read than raw text keeps recurring across separate research groups.","[\"visual text compression\",\"long-context ai\",\"arxiv research\",\"llm efficiency\"]","2026-09-30T04:00:00.000Z","2026-09-30T14:11:12.000Z","2026-09-30T14:11:17.112Z","published",null,[],"ai",[26,27,28,29],"visual text compression","long-context ai","arxiv research","llm efficiency",[31],{"name":32,"url":33},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.36651",0,{"sections":36},[37,40,44,48,53,58,63,68,73,77,82,87,92,97],{"name":38,"slug":24,"count":39,"latest_published_at":18},"AI",5135,{"name":41,"slug":42,"count":43,"latest_published_at":18},"Security","security",788,{"name":45,"slug":46,"count":47,"latest_published_at":18},"Policy","policy",417,{"name":49,"slug":50,"count":51,"latest_published_at":52},"Deals","deals",284,"2026-09-29T21:00:00.000Z",{"name":54,"slug":55,"count":56,"latest_published_at":57},"Hardware","hardware",194,"2026-09-29T13:16:04.000Z",{"name":59,"slug":60,"count":61,"latest_published_at":62},"Science","science",154,"2026-09-28T13:19:18.000Z",{"name":64,"slug":65,"count":66,"latest_published_at":67},"Consumer Tech","consumer-tech",142,"2026-09-29T18:38:03.000Z",{"name":69,"slug":70,"count":71,"latest_published_at":72},"Software","software",91,"2026-09-25T20:55:00.000Z",{"name":74,"slug":75,"count":76,"latest_published_at":18},"Dev Tools","dev-tools",90,{"name":78,"slug":79,"count":80,"latest_published_at":81},"Startups","startups",83,"2026-09-29T21:51:36.000Z",{"name":83,"slug":84,"count":85,"latest_published_at":86},"General","general",49,"2026-09-28T16:44:57.000Z",{"name":88,"slug":89,"count":90,"latest_published_at":91},"Gaming","gaming",48,"2026-09-25T18:35:21.000Z",{"name":93,"slug":94,"count":95,"latest_published_at":96},"Reviews","reviews",31,"2026-09-28T14:31:34.000Z",{"name":98,"slug":99,"count":100,"latest_published_at":101},"How-To","how-to",6,"2026-06-16T09:00:00.000Z"]